In nabla9's example, S would be the "vector discrepancy". If you want to minimize the discrepancy of s and X (NB. s and X, not S and X, I think this was a mista

In nabla9's example, S would be the "vector discrepancy". If you want to minimize the discrepancy of s and X (NB. s and X, not S and X, I think this was a mistake), you want this vector to be as small as possible. In other words, you want the magnitude of the S vector to be minimal. This magnitude is given by the norm of the vector, sqrt(sum((S_i)^2)), which will reach a minimum when sum((S_i)^2) reaches a minimum.